Performance Studies using PSTSWM

SGI Origin2000 SWAP Performance

(ordered swap of 8KB message using SHMEM)

ordered swap using nonblocking send

Data Statistics unidirectional bandwidth estimated latency model error (peak MByte/sec) (usec/msg) (max. rel. error) 1 iter. 63.60 6.15 11.5% 10 iter. 311.25 5.98 25.6% 1 iter. w/overlap 58.68 6.54 11.5% 10 iter. w/overlap 295.53 5.14 24.7%

ordered swap using nonblocking receive

Data Statistics unidirectional bandwidth estimated latency model error (peak MByte/sec) (usec/msg) (max. rel. error) 1 iter. 76.99 6.22 12.5% 10 iter. 302.96 6.09 37.3% 1 iter. w/overlap 97.06 6.13 51.0% 10 iter. w/overlap 407.97 3.52 20.2%

synchronous

Data Statistics unidirectional bandwidth estimated latency model error (peak MByte/sec) (usec/msg) (max. rel. error) 1 iter. 79.69 5.78 10.5% 10 iter. 353.71 5.66 24.4%

DISCUSSION

• An Origin2000 processing node consists of one or two processors. If two processors are in a node, they share the node's network interface hardware. It can be slower to exchange data between processors in the same node than between processors on different nodes (due to contention for shared resources?). For this reason, our SWAP experiments used logical processors 0 and 2 in a 4 processor partition.
• The experiments using multiple iterations show the effect of the improved memory locality intrinsic to the iteration logic, giving significantly better performance for the smaller message sizes. This is unlikely to be indicative of performance in real applications.

---

Patrick H. Worley / ( worleyph@ornl.gov)
Last Modified Monday, 15-Jul-2002 10:12:19 EDT.
86312 accesses since 1/2/96.